Reassessing seasonal sea ice predictability of the Pacific-Arctic sector using a Markov model

In this study, a regional linear Markov model is developed to assess seasonal sea ice predictability in the Pacific-Arctic sector. Unlike an earlier pan-Arctic Markov model that was developed with one set of variables for all seasons, the regional model consists of four seasonal modules with different sets of predictor variables, accommodating seasonally varying driving processes. A series of sensitivity tests are performed to evaluate the predictive skill in cross-validated experiments and to determine the best model configuration for each season. The prediction skill, as measured by the sea ice concentration (SIC) anomaly correlation coefficient (ACC) between predictions and observations, increased by 32 % in the Bering Sea and 18 % in the Sea of Okhotsk relative to the pan-Arctic model. The regional Markov model's skill is also superior to the skill of an anomaly persistence forecast. SIC trends significantly contribute to the model skill. However, the model retains skill for detrended sea ice extent predictions for up to 7-month lead times in the Bering Sea and the Sea of Okhotsk. We find that subsurface ocean heat content (OHC) provides a crucial source of prediction skill in all seasons, especially in the cold season, and adding sea ice thickness (SIT) to the regional Markov model has a substantial contribution to the prediction skill in the warm season but a negative contribution in the cold season. The regional model can also capture the seasonal reemergence of predictability, which is missing in the pan-Arctic model.